Mellum2.1 review, pricing and limits

JetBrains released Mellum2.1 on 8 October 2026 as a text-only model with a 131,072-token context window. It suits teams that want a small, self-hosted worker for coding sub-agents on hardware they control, rather than a frontier-sized model that needs an API contract.

Mellum2.1 is JetBrains' open-weights reasoning model with 12B total and 2.5B active parameters. JetBrains reports 47.0% on SWE-bench Verified, up from near zero for Mellum2, after reinforcement learning in real repositories. It runs on your own hardware, so code never leaves your infrastructure.

Provider: JetBrains · Family: Mellum

More about JetBrains on HokAI

Context window: 131,072 tokens

Input modalities: text, tool-calls · Output: text, tool-calls

About Mellum2.1

Mellum2.1 is an open-weights reasoning model from JetBrains, released on 8 October 2026 as the successor to Mellum2, which the company open-sourced in June. The architecture did not change: it is a mixture-of-experts transformer with 12B total parameters, of which 2.5B are active for each token, built from 28 layers with 64 experts and 8 of them activated per token. The public checkpoint is called Mellum2.1-12B-A2.5B-Thinking, and JetBrains describes it as a thinking model for working inside a repository, running commands and calling tools.

What changed is post-training. JetBrains says reinforcement learning grew from a short final stage into the main part of training, with millions of sandboxed runs across thousands of environments. For software engineering the model trained inside real repositories with a shell and file-editing tools and was rewarded when the tests passed. The payoff shows on the agentic rows of JetBrains' own table: SWE-bench Verified went from 2.0 for Mellum2 to 47.0, Terminal-Bench 2.1 reached 17.4 and SWE-bench Pro reached 28.0. It also posts 82.0 on LiveCodeBench v6 and 64.6 on GPQA Diamond. Every figure is self-reported by JetBrains, run through one pipeline for all four models in the table.

The context length is 131,072 tokens, and three of every four layers use a sliding attention window of 1,024 tokens, which keeps memory growth in check on long sessions. Weights ship as bfloat16 safetensors under Apache 2.0, and JetBrains documents serving with vLLM, including a tool-calling mode. A GGUF repository is already listed in the Hugging Face collection for llama.cpp, Ollama and LM Studio, while the blog post says a multi-token-prediction head for vLLM is still to come. On speed, JetBrains reports that under heavy load on one H200 GPU it serves almost twice as many tokens as Qwen3.5-9B.

It is a good fit for a cheap sub-agent that does the bulk of the reading, editing and test-running in a private repository, because the weights run on your own infrastructure and code never leaves it. Open coding agents such as Cline, OpenHands and OpenCode are the kind of client to pair it with; check each one's custom-endpoint settings first. It is the wrong pick when you need frontier-level reasoning, images or audio. Much larger open models such as GLM-5.3, at about 753B total parameters, trade far heavier hardware for more capability, and Gemma 4 12B from Google DeepMind is the multimodal small-model option. Other open-weights models on HokAI to compare include Kimi K2.7 Code, DeepSeek V4 Flash and GLM-5.3 Flash. The weights are distributed through Hugging Face.

Pricing

Mellum2.1 is free to self-host under its open licence. JetBrains does not sell hosted inference for it in the launch post or the model card, so there is no per-token price to quote; your cost is the GPU time you provide.

Key Features

• Sparse mixture of experts: Many small experts with only a few active per token keep compute far below the total parameter count.
• RL in real repositories: Trained with a shell and file-editing tools across millions of sandboxed runs, rewarded when tests pass.
• Long context window: Sliding-window attention on most layers limits memory growth on long sessions.
• Open weights: Download from Hugging Face and serve with vLLM or a GGUF runtime on your own hardware, under an open licence.

Pros

• Fully open licence means no vendor lock-in and no data leaving your servers.
• Small active footprint makes it practical as a high-volume sub-agent.
• JetBrains publishes the full comparison table, including rows where rivals win.

Cons

• Self-reported benchmarks only, run by the vendor on its own pipeline.
• Text only, so it cannot read screenshots or diagrams.
• You carry the hosting, scaling and update burden yourself.

Benchmarks

• GPQA Diamond: 64.6% vendor-reported · 20 Oct 2026 — PhD-level science questions that are hard to search for, % correct.
• SWE-bench Pro: 28% vendor-reported · 20 Oct 2026 — Harder, longer real-repository coding tasks, % solved.
• LiveCodeBench v6: 82% vendor-reported · 20 Oct 2026 — Recently published coding problems, % solved.
• SWE-bench Verified: 47% vendor-reported · 20 Oct 2026 — Real GitHub issues fixed end to end, % solved.
• Terminal-Bench 2.1: 17.4% vendor-reported · 20 Oct 2026 — Multi-step tasks completed in a real command line, % solved.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Mellum2.1 Official Page